MV2MAE: Self-Supervised Video Pre-Training with Motion-Aware Multi-View Masked Autoencoders

Abstract

Videos captured from multiple viewpoints can help in perceiving the 3D structure of the world and benefit computer vision tasks such as action recognition, tracking, etc. In this paper, we present MV2MAE, a method for self-supervised learning from synchronized multi-view videos, built on the masked autoencoder framework. We introduce two key enhancements to better exploit multi-view video data. First, we design a cross-view reconstruction task that leverages a cross-attention-based decoder to reconstruct a target viewpoint video from source view. This helps in effectively injecting geometric information and yielding representations robust to viewpoint changes. Second, we introduce a controllable motion-weighted reconstruction loss which emphasizes dynamic regions and mitigates trivial reconstruction of static backgrounds. This improves temporal modeling and encourages learning more meaningful representations across views. MV2MAE achieves state-of-the-art results on the NTU-60, NTU-120 and ETRI datasets among self-supervised approaches. In the more practical transfer learning setting, it delivers consistent gains of +2.0 -- 8.5% on NUCLA, PKU-MMD-II and ROCOG-v2 datasets, demonstrating the robustness and generalizability of our approach. Code: https://github.com/kshah33/mv2mae